Following the model run, the model results of the entire agent population are extrapolated at sectoral level with the aid of a weighting process. SWISSland calculates sec-toral parameters via an extrapolation algorithm. Various sectoral output indicators are of interest: product quan-tities and prices, land use and labour trends, income trend according to the Economic Accounts for Agricul-ture, sectoral input and output factors for calculating environmental impacts, and important key structural figures such as number of farms, sizes and types of farm, or number of farms switching farming system. Zimmer-
An essential aim of the SWISSland model is to forecast structural change in Swiss agriculture. Structural change is a result of decisions taken at individual-farm level. A multi-agent model depicting all of the approx. 50 000 farms in Switzerland would be hard to implement, howe-ver, since the necessary data are not available from all of the farms, and the high volume of data and long compu-ter run times would make application extremely difficult.
For these reasons, SWISSland only depicts a sample of all farms, which makes an extrapolation necessary for secto-ral statements.
Our objective is to progress from the individual-farm results of a model to a regional, sectoral or structural level. Hence, we need to develop a method that reflects
Table 3.8: Possible representation errors for the entire region in various sector models
Model Type Aggregation
Error Sampling Error Regional farm Entire region as virtual farm +++ (-)
Farm types Groups of average farms ++ +
Farm sample
Representative selection (-) ++
Non-representative
selection (-) +++
Complete coverage of the
region (-) (-)
+++ = strong probability; ++ = some probability;
+ = weak probability; (-) = no probability (3-20)
both the reactions of the farm model and the official numbers from the statistics as faithfully as possible. This method must be based on attributes that are available both from the individual farms and from the basic popu-lation. In Switzerland, only structural attributes such as surface areas of cultivated crops and livestock numbers are systematically collected from all farms. Economic figures are only available from the FADN farms. A compa-rison of important structural attributes for the basic population and for the farm sample indicates that some attributes are strongly under- or over-represented. The percentage of small-sized farms and the surface areas of permanent crops are much lower in the sample, whilst the dairy-cow population is higher. The average area of an individual farm is also higher in the sample (Zimmer-mann et al., 2015).
The aim of the extrapolation is to apply the results of the sample-based model to the appropriate basic population, using specific methods. For this, a suitable weight is gene-rally sought for every micro-unit (individual farm) of a micro-database (sample). Consequently, the sum of attri-butes formed in each case with these weights should cor-respond as closely as possible across all micro-units to the given data of the basic population. In the literature, diffe-rent objective/distance functions are utilised for this, such as generalised least squares and minimum information loss. In this optimisation process, the extrapolation factors can be determined by minimising the deviations either from initial extrapolation factors or from the statistical characteristics.
In multi-agent models where relationships such as land trade exist between the agents, however, the allocation of individual-farm extrapolation factors can lead to
inconsis-tencies in the extrapolation: Land trade between farms to which different extrapolation factors are assigned leads to a change in the stipulated total area (Figure 3.6, column 2).
This change in area can be corrected by a corresponding adjustment of the farm extrapolation factors, i.e. the fac-tors are adjusted so as to leave the extrapolated total area unchanged (Figure 3.6, right). At the same time, however, such a correction has repercussions for the extrapolation of all remaining farm attributes, whose extrapolated values would change not because of the model calculati-ons, but simply because of the correction of the extrapola-tion factors. To prevent such inconsistencies, the proxy of the farms in the agent population could be adjusted beforehand to the proxy in the basic population, based on the initialisation method of Happe (2004). Unlike in Happe (2004), however, the number of certain farms would not be multiplied until the entire region was covered. Instead, farms from under-represented farm groups would be mul-tiplied and some from over-represented groups would be removed from the agent population, if necessary. The goal of this adjustment is for a similar percentage of all essen-tial attributes to be represented, allowing the results of model calculations to be extrapolated to the basic popula-tion with a general, fixed factor.
The lessons drawn from the analysis described in Zimmer-mann et al. (2015) can be grouped into two categories:
those drawn for improving the SWISSland model, and those drawn from a purely methodological perspective.
Starting with the second category, applying optimisation models in such a way as to allow comparison with obser-ved developments has been shown to be potentially hel-pful. This is probably the only way to normatively compare different methodological options that are all theoretically plausible. Furthermore, a validation process in which
diffe-Figure 3.6: Aggregation errors of land trade with farm-specific extrapolation factors.
rent options are analysed has a positive influence on the reliability of the model results (see also Chapter 7). On the other hand, it must be conceded that methods that have worked well in the past will not necessarily work well in future.
As is probably true of most FADN networks, the Swiss FADN does not constitute a fully representative sample of Swiss agriculture – a fact that should always be borne in mind when it is used as the main data source for a forecas-ting model. The SWISSland model has been shown to yield the best results when permitted to use a certain group of farms more than other groups. An adjustment of the sam-ple by the multiplication of under-represented farms and if necessary the removal of over-represented ones showed a better alignment with the observed trends, and prevents inconsistencies arising from relationships between farm agents assigned different extrapolation factors. On the other hand, an optimisation of individual-farm extrapola-tion factors could help to enhance alignment with the population as a whole. Furthermore, research is needed to determine which method would be the most appropriate in cases of greater changes in economic or political condi-tions within the time period under consideration. As every model differs in terms of its structure, underlying data and objectives, the most suitable method for extrapolation to the sector must probably be determined separately for each model. The sensitivity of the model to the aggrega-tion approach as well as the lack of any clear ”winner” in terms of which approach should be taken suggests that
”agent-based models with aggregated agents” is either a fundamentally flawed concept (since e.g. individual-agent interactions such as land exchange are problematic to represent), or at the very least an area requiring a vast amount of work.
where i depicts oilseed category, t indicates time, and Con-scruit is a measure that captures the past interaction bet-ween the crush margin and crushing. In this specification, the demand for processing increases along with increases in the processing/crushing margins, and vice versa. In other words, as the processing margin increases, there will be a greater demand for oilseeds for processing, resulting in a gradual rise in oilseeds prices. is the crush elasticity with respect to own price i and cross-price j of oilseeds, whilst is a partial adjustment parameter. The crushing margin MGpit is specified as a function of the extraction rate of crush products, the prices of crush products (meal and oil), and the consumer prices for oilseeds.
Processing supply is defined as processing demand multi-plied by the respective extraction factor. Production of oil-seed products i at year t (PRDit) is determined by the quan-tity of ith oilseed crushed and by an exogenous extraction rate as follows: